Avoid pulling in libcu++ with cooperative_groups.h

[eyalroz]

When <cooperative_groups.h> is included, it pulls in a chunk of libcu++ (blech). Beginning with CUDA 13.0, that library requires C++17 - and thus, so does <cooperative_groups.h>. But - we're a C++11 library. Luckily, there's a safety valve in the form of _CG_LIMIT_INCLUDED_DEPENDENCIES, a preprocessor macro which, when defined, prevents the pulling in of libcu++ constructs. So, let's use it when we use cooperative groups in example programs.